The same document, moved
Twenty documents supplied, exactly one of which holds the answer. Only its position in the pile changes. Everything else is held constant.
The middle bar sits below the dashed line. On this test, handing the system the right document in the worst position left it worse off than handing it no documents at all.
Source: Liu et al., "Lost in the Middle: How Language Models Use Long Contexts", arXiv:2307.03172 (2023), published in TACL (2024). Figures shown are the 20-document setting for GPT-3.5-Turbo. These are 2023 systems and should not be read as current performance.